vllm-project/recipes

Common recipes to run vLLM

vLLM is an engine for running large AI models. This project helps people answer a practical question: How do I run model X on hardware Y for task Z? — for example, how to run DeepSeek-V3 on NVIDIA GPUs for chat, or Qwen3 for image recognition.

This repository is the central hub for those instructions. Each instruction is called a recipe — a structured file describing the best way to run a specific model, on specific hardware, for a specific use case.

How it works

The recipes themselves are written and maintained by the teams that know each model best (for example, DeepSeek's own team maintains the DeepSeek recipe). Those recipes live in other repositories, not here. This repo's job is to:

  1. Pull in all those recipe files from their home repos
  2. Merge them together
  3. Validate them
  4. Build the public website recipes.vllm.ai and its JSON API

Think of it like a newspaper editor: the articles are written by specialists in different cities, and this repository is the newsroom that collects, checks, and prints them.

Automatic updates

A background process checks every 20 minutes whether any recipe source has changed. If yes, it validates the new content, then updates the site automatically. If validation fails, the site keeps serving the last good version — nothing breaks — and an issue is opened on GitHub to flag the problem.

Provenance and trust

Every recipe page includes an "Edit recipe" link pointing to the repository that actually owns that file. Every piece of recipe data carries metadata identifying the exact repo, commit, and file path it came from — so you can always trace a command back to its original source.

Who contributes what

This project separates responsibilities:

  • Model recipes (like "How to run DeepSeek-V3") are owned by the vLLM project and the vLLM-Omni project
  • The schema, hardware profiles, serving strategies, and platform definitions are owned here
  • The website build, JSON API, and command synthesis are owned here

Local development

Developers can preview changes locally by pointing this repo at a local checkout of a source repository (via a sources.local.yaml file), then running a few commands to sync, validate, and assemble the content.

Legacy content

There are also older Markdown guide files (in folders like DeepSeek/, Qwen/, etc.) — these are from a previous documentation site, kept as a reference while the project transitions to the newer YAML-based recipe format.

License

Apache License 2.0.

Related

  • Project
    sybil-solutions/codex-shim

    codex‑shim is a Python/aiohttp local server that mimics the OpenAI Responses API used by Codex Desktop. It lets you add arbitrary BYOK LLMs (OpenAI, Anthropic, Ollama, etc.) to Codex’s model picker without rebuilding the app, preserving full Codex features (function calls, streaming, images). The shim reads a `models.json` config, generates a Codex‑compatible catalog, runs on 127.0.0.1:8765, and routes each request to the chosen upstream, optionally falling back to ChatGPT’s Codex model. It works on macOS, Linux, WSL, and Windows, with optional macOS picker patch for hidden‑model filtering.

  • Project
    onnx/onnx-mlir

    A compiler toolset that transforms ONNX neural network models into optimized, standalone executable code or libraries using LLVM/MLIR technology.

  • Project
    Xilinx/finn

    An experimental framework from AMD Research that enables fast, scalable inference of quantized neural networks on FPGAs by generating customized dataflow architectures.

  • Project
    stan-dev/rstan

    RStan is the R interface to Stan, a probabilistic programming language for Bayesian inference. It lets R users write Stan models, compile them, and run state‑of‑the‑art MCMC or variational inference, returning results as R objects that integrate with standard diagnostic and plotting tools.